Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning...
Transcript of Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning...
![Page 1: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/1.jpg)
Self-paced Curriculum Learning
Lu Jiang 1, Deyu Meng2, Qian Zhao 2, Shiguang Shan 3, Alexander G. Hauptmann 1
1School of Computer Science, Carnegie Mellon University
2School of Mathematics and Statistics, Xi'an Jiaotong University
3Institute of Computing Technology, Chinese Academy of Sciences
1 AAAI 2015
![Page 2: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/2.jpg)
People
• CMU Informedia Team
Qian Zhao Deyu Meng Alexander Hauptmann Shiguang Shan
![Page 3: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/3.jpg)
Outline
Motivation
Background Knowledge
Self-paced Curriculum Learning
Experiments
Conclusions
3
![Page 4: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/4.jpg)
Motivations
• Noisy
• Highly unbalanced data
• Training non-convex models?
4
Many algorithms have been proposed. One solution is biologically inspired: what we will do if we are asked to learn something from the big data.
![Page 5: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/5.jpg)
Curriculum Learning and Self-paced Learning
• Curriculum Learning (Bengio et al. 2009) or self-paced learning (Kumar et al 2010) represents a recently proposed learning paradigm that is inspired by the learning process of humans and animals.
• The samples are not learned randomly but organized in a meaningful order which illustrates from easy to gradually more complex examples.
• Curriculum: a sequence of gradually learned samples.
Y. Bengio, J. Louradour, R. Collobert, and J. Weston. Curriculum learning. In ICML, 2009. M. P. Kumar, B. Packer, and D. Koller. Self-paced learning for latent variable models. In NIPS, pages 1189–1197, 2010. 5
![Page 6: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/6.jpg)
Curriculum Learning and Self-paced Learning
• Samples are organized in a meaning order (curriculum).
• Learning is conducted iteratively.
• Models are becoming increasingly complex.
Age
6
![Page 7: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/7.jpg)
Curriculum Learning and Self-paced Learning
“bus” samples to learn earlier “bus” samples to learn later
Age
7
• Samples are organized in a meaning order (curriculum).
• Learning is conducted iteratively.
• Models are becoming increasingly complex.
*The above of real examples in the TRECVID SIN dataset (http://trecvid.nist.gov/).
![Page 8: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/8.jpg)
Samples of “Dog” to learn earlier.
Samples of “Dog” to learn later.
Easy and Complex samples in Google Image Search
8 In Big data, we see a lot more examples like this.
![Page 9: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/9.jpg)
Outline
Motivation
Background Knowledge
Self-paced Curriculum Learning
Experiments
Conclusions
9
![Page 10: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/10.jpg)
Curriculum Learning
• Curriculum Learning (CL): assign learning priorities to training samples, according to prior knowledge or heuristics about specific problems.
• Teaching a robot: leverage human curriculum.
10
Y. Bengio, J. Louradour, R. Collobert, and J. Weston. Curriculum learning. In ICML, 2009.
Khan, F.; Zhu, X.; and Mutlu, B. 2011. How do humans teach: On curriculum learning and teaching dimension. In NIPS.
![Page 11: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/11.jpg)
Curriculum Learning
• Curriculum Learning (CL): assign learning priorities to training samples, according to prior knowledge or heuristics about specific problems.
• Teaching a robot: leverage human curriculum.
• Parsing in Natural Language Processing (NLP):
– From shorter sentences to longer sentence.
11
Spitkovsky, V. I.; Alshawi, H.; and Jurafsky, D. 2009. Baby
steps: How less is more in unsupervised dependency parsing. In NIPS
![Page 12: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/12.jpg)
Self-paced Learning
• Self-paced Learning (SPL): the curriculum is determined by the learned models.
• Solving a joint optimization problem of the learning objective with the curriculum (a sequence of gradually added samples).
– From smaller loss to the already learned model.
– to larger loss to the already learned model.
12
M. P. Kumar, B. Packer, and D. Koller. Self-paced learning for latent variable models. In NIPS, pages 1189–1197, 2010. Jiang, L.; Meng, D.; Yu, S.-I.; Lan, Z.; Shan, S.; and Hauptmann, A. G. 2014b. Self-paced learning with diversity. In NIPS.
![Page 13: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/13.jpg)
Curriculum Learning versus Self-paced Learning
13
• Pros – Flexible to incorporate prior
knowledge/heuristics.
• Cons – Curriculum is determined
beforehand which may not be consistent with dynamically learned modles.
Curriculum Learning (CL)
• Pros – Learn consistent models.
– Concise optimization problem.
• Cons – Cannot use prior knowledge.
– Random starting values (can be sensitive to the performance).
Self-paced Learning (SPL)
Difficult to judge which one is better in practice.
![Page 14: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/14.jpg)
Curriculum Learning versus Self-paced Learning
14
Curriculum Learning (CL)
Self-paced Learning (SPL)
instructor-driven student-driven
Difficult to judge which one is better in practice.
![Page 15: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/15.jpg)
Self-paced Curriculum Learning
15
Curriculum Learning (CL)
Self-paced Learning (SPL)
Self-paced Curriculum Learning (SPCL) instructor-driven
student-driven
instructor-student-collaborative
Unified in a single framework: SPCL
![Page 16: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/16.jpg)
Outline
Motivation
Background Knowledge
Self-paced Curriculum Learning
Experiments
Conclusions
16
![Page 17: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/17.jpg)
Self-paced Curriculum Learning
• Formulated as an optimization problem (based on SPL). Consider a binary classification problem:
17
Off-the-shell model (SVM, deep neural networks etc.)
Weight vectors in self-paced learning
Prior knowledge in curriculum learning
![Page 18: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/18.jpg)
Self-paced Curriculum Learning
How to solve (alternative search):
– Fixing v and optimize model parameters w .
– Fixing w and optimize weight variables v .
– Increase the model age to train a more complex model.
18
smaller loss bigger loss
Training a model
M. P. Kumar, B. Packer, and D. Koller. Self-paced learning for latent variable models. In NIPS, pages 1189–1197, 2010.
![Page 19: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/19.jpg)
Self-paced Curriculum Learning
How to solve (alternative search):
– Fixing v and optimize model parameters w .
– Fixing w and optimize weight variables v .
– Increase the model age to train a more complex model.
19
smaller loss bigger loss
Recalculating the loss and select more examples.
![Page 20: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/20.jpg)
Self-paced Curriculum Learning
How to solve (alternative search):
– Fixing v and optimize model parameters w .
– Fixing w and optimize weight variables v .
– Increase the model age to train a more complex model.
20
smaller loss bigger loss
Increase the model age to include more examples
![Page 21: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/21.jpg)
Self-paced Curriculum Learning
• Formulated as an optimization problem (based on SPL):
21
![Page 22: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/22.jpg)
Self-paced Curriculum Learning
• Formulated as an optimization problem (based on SPL):
• Novelty: when optimizing with the fixed : – Encode heuristics/prior knowledge in the feasible region :
• E.g. v1 learned before v3, v2 before v3
– Represent the regularizer to present different learning scheme. Apply different regularizer to different problems: • Start from easy to complex examples?
• From easy and diverse to complex examples?
• Even from complex to easy (for very smart learner/student for example)?
22
![Page 23: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/23.jpg)
Self-paced Curriculum Learning
• Formulated as an optimization problem (based on SPL):
• When optimizing with the fixed : – Encode heuristics/prior knowledge in the feasible region :
• E.g. v1 learned before v3, v2 before v3
– Represent the regularizer to present different learning scheme. Apply different regularizers to different problems: • Start from easy to complex examples?
• From diverse to complex examples?
• Even from complex to easy (for very smart learner/student for example)?
23
![Page 24: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/24.jpg)
Outline
Motivation
Background Knowledge
Self-paced Curriculum Learning
Experiments
Conclusions
24
![Page 25: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/25.jpg)
Experiments
• Matrix factorization:
• Content-based video retrieval:
25
![Page 26: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/26.jpg)
Experiments
26
RMSE (Root Mean Square Error)
Lower -> better
MAP(Mean Average Precision)
Higher -> better
Incorporating prior knowledge into statistical learning tends to be instrumental.
![Page 27: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/27.jpg)
Outline
Motivation
Background Knowledge
Self-paced Curriculum Learning
Experiments
Conclusions
27
![Page 28: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/28.jpg)
Conclusions
Take home messages:
– Proposed a novel learning framework that unifies the existing curriculum learning and self-paced learning paradigms.
– SPCL is general and has pluggable components:
• Off-the-shell model Student
• Regularizers Learning schemes
• Feasible region Prior knowledge
– Observed benefits for the non-convex problems and the problems with noisy and unbalanced data.
28
![Page 29: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/29.jpg)
29
![Page 30: Self-paced Curriculum Learninglujiang/resources/SPCL_AAAL15.pdf · Self-paced Curriculum Learning Lu Jiang 1, Deyu 3Meng2, Qian Zhao2, Shiguang Shan , Alexander G. Hauptmann 1 1School](https://reader033.fdocuments.net/reader033/viewer/2022041515/5e2ab4e0704075271f5fb29e/html5/thumbnails/30.jpg)
Self-paced Curriculum Learning
• Self-paced curriculum (SPCL) Learning unified curriculum learning (CL) and self-paced learning (SPL) into a universal framework.
30